whisper-talk

System-wide voice dictation for Linux with Whisper and Parakeet-v3 backends.

Overview

whisper-talk is a lightweight, efficient voice dictation daemon for Linux that provides system-wide, local speech-to-text with either Whisper or NVIDIA Parakeet-v3. Press a global hotkey, speak, and have the transcribed text copied to your clipboard so you can paste it into any application.

Key Features

• Privacy-focused - Local transcription only, no cloud services required
• Two transcription backends - Whisper for broad model and translation support, or Parakeet-v3 for fast multilingual transcription
• CPU and GPU acceleration - CPU by default, NVIDIA CUDA and AMD ROCm support
• Recording modes - Toggle, Push-to-Talk, and Auto (hybrid) modes
• OpenAI compatible whisper endpoint - Use it in other applications as well

Backend and GPU Support

Backend	CPU	NVIDIA CUDA	AMD ROCm	Translation to English
Whisper	Yes	--features cuda	--features hipblas	Yes
Parakeet-v3	Yes	--features cuda	Yes, with a ROCm-enabled ONNX Runtime	No

Parakeet-v3 requires a compatible ONNX Runtime 1.20.1 shared library for CPU, CUDA, or ROCm. See INSTALL.md for platform-specific runtime and driver setup.

How It Works

1. Press your configured global shortcut (default: SUPER+ALT+D)
2. whisper-talk starts recording from your microphone
3. Speak your text
4. Press the shortcut again
5. Audio is transcribed locally using your selected Whisper or Parakeet-v3 backend
6. Transcribed text is injected into clipboard

Quick Start

Prerequisites

• Linux system (Fedora, Ubuntu, Arch, or derivative)
• Rust 1.70+ for building from source
• Microphone input device
• ydotool for clipboard injection
• ONNX Runtime 1.20.1 when using Parakeet-v3
• CUDA or ROCm libraries only when building for the corresponding GPU backend

Installation

# Clone the repository
git clone https://github.com/aljazceru/whisper-talk.git
cd whisper-talk

# Install base system dependencies (Fedora)
sudo dnf install alsa-lib-devel pulseaudio-libs-devel ydotool

# Or for Ubuntu/Debian
sudo apt install libasound2-dev libpulse-dev ydotool

# Build and install for CPU
cargo install --path .

# Or build and install with NVIDIA CUDA support for Whisper and Parakeet-v3
cargo install --path . --features cuda

# Or build and install with AMD ROCm support for Whisper
# Parakeet-v3 uses a separately installed ROCm-enabled ONNX Runtime.
cargo install --path . --features hipblas

# Run setup wizard
whisper-talk setup

Initial Setup

The interactive setup wizard guides you through:

1. Selecting Whisper or Parakeet-v3 as the transcription backend
2. Choosing a Whisper model size, or downloading the Parakeet-v3 model
3. Selecting your audio input device
4. Setting up user permissions
5. Optionally installing systemd service

whisper-talk setup

You can also download and select a backend manually:

# Whisper (default)
whisper-talk model download base
whisper-talk config set transcription.backend whisper
whisper-talk config set transcription.model base

# Parakeet-v3 (install ONNX Runtime 1.20.1 first)
whisper-talk model download parakeet-v3
whisper-talk config set transcription.backend parakeet_v3
whisper-talk config set transcription.model parakeet-v3

Basic Usage

# Start the daemon
whisper-talk daemon

# Or run in background with systemd
systemctl --user start whisper-talk

# Press SUPER+ALT+D to start/stop recording

OpenAI-Compatible API

Run the daemon with either:

whisper-talk daemon --api-bind 127.0.0.1:11434

Or configure it in the config file (see below):

{
  "api_bind": "127.0.0.1:11434"
}

If neither is set, API is disabled and existing local hotkey behavior is unchanged.

The API serves OpenAI-compatible routes under /v1 (for example /v1/audio/transcriptions).

Then request transcription with OpenAI-compatible multipart format:

curl -X POST \
  -F file=@/path/to/audio.wav \
  -F model=whisper-1 \
  -F language=en \
  -F response_format=json \
  http://127.0.0.1:11434/v1/audio/transcriptions

Response:

{ "text": "..." }

Supported uploads include WAV by default and common formats like MP3/OGG/FLAC when detected. Response format is JSON by default and text when response_format=text.

Whisper also supports translation requests to English:

curl -X POST \
  -F file=@/path/to/audio.mp3 \
  -F model=whisper-1 \
  -F prompt="Translate the audio to English" \
  http://127.0.0.1:11434/v1/audio/translations

The Parakeet-v3 backend supports transcription only and returns an unsupported-operation error for this endpoint.

Configuration

Configuration is stored in ~/.config/whisper-talk/config.json:

{
  "shortcuts": {
    "primary_shortcut": "SUPER+ALT+d",
    "recording_mode": "toggle",
    "grab_keys": false,
    "auto_mode_threshold_ms": 400
  },
  "audio": {
    "device_id": null,
    "device_name": null,
    "mute_detection": true,
    "zero_volume_threshold": 5e-7
  },
  "transcription": {
    "backend": "whisper",
    "model": "base",
    "threads": 4,
    "language": null,
    "word_overrides": {},
    "whisper_prompt": "Transcribe with proper capitalization, including sentence beginnings, proper nouns, titles, and standard English capitalization rules."
  },
  "injection": {
    "paste_mode": "ctrl_shift",
    "auto_submit": false,
    "clipboard_behavior": false,
    "clipboard_clear_delay": 5.0
  },
  "feedback": {
    "mic_osd_enabled": true,
    "audio_feedback": true,
    "master_volume": 1.0,
    "start_sound_volume": 1.0,
    "stop_sound_volume": 1.0,
    "error_sound_volume": 1.0
  },
  "api_bind": "127.0.0.1:11434"
}

System Integration

systemd Service

Enable whisper-talk to start automatically on login:

# Generate and install service file
whisper-talk systemd install

# Enable and start
systemctl --user enable whisper-talk
systemctl --user start whisper-talk

# Check status
systemctl --user status whisper-talk

Troubleshooting

No audio input detected

# Check if microphone is detected by system
pactl list sources short

# Or with ALSA
arecord -l

# Verify audio permissions
groups $USER  # Should include 'audio'

Text not being typed

# Test ydotool
echo "test" | ydotool type -

# Check ydotool permissions
groups $USER  # Should include 'ydotool'

# Start ydotool daemon (if needed)
ydotool --version

Global shortcut not working

# Check input device permissions
groups $USER  # Should include 'input'

# Check device access
ls -l /dev/input/event*

Transcription errors

# Check model exists
whisper-talk model list

# Validate configuration
whisper-talk validate

# Check logs
journalctl --user -u whisper-talk -f

Performance issues

• Reduce model size (use tiny or base instead of medium or large-v3)
• Build with --features cuda for NVIDIA acceleration with Whisper or Parakeet-v3
• Build with --features hipblas for AMD ROCm acceleration with Whisper
• Install a ROCm-enabled ONNX Runtime 1.20.1 for AMD acceleration with Parakeet-v3

Inspiration

Project was inspired by hyprwhspr.